The Education-Health Nexus: a Meta-Analysis

HEDG HEALTH, ECONOMETRICS AND DATA GROUP

WP 20/01

The Education-Health Nexus: A Meta-Analysis

Xindong Xue

January 2020

http://www.york.ac.uk/economics/postgrad/herc/hedg/wps/

The Education-Health Nexus: A Meta-Analysis

Xindong Xue1

School of Public Administration, Zhongnan University of Economics & Law, China

December 6, 2019

1 Correspondence: No.182, Nanhu Avenue, East Lake Hi-tech Development Zone, Wuhan 430073, China. Tel.: +86 27 8838 7901; fax: +86 27 8838 6936. E-mail address: [email protected].

The Education‐Health Nexus: A Meta‐Analysis

Xindong Xue

Abstract

Does education cause a better health? No consensus answer to this question has yet emerged. In this paper, I perform a meta-analysis of the extensive literature on the health effects of education. The final sample identifies 105 studies with 4,671 estimates. Overall, the health effect of education is not economically meaningful, although statistically significant. There is severe publication bias favoring the positive effect of education on health. Studies that do not control for endogeneity are prone to exaggerate the estimated effect. In addition, the effect becomes weaker for more recent studies. The results suggest that education may not be a feasible policy option for promoting population health.

JEL CODES: B49, C49, I10, I20, I31 KEYWORDS: Education, Health, Human capital, Meta-analysis, Research synthesis

1.Introduction

Since the seminal works of Schultz (1961), Becker (1964) and Grossman (1972), social

scientists have generally agreed that education and health play a fundamental role in economic

development and well-being. In 2015, United Nations listed education and health as important

sustainable development goals by 2030 (United Nations, 2015). Given their intrinsic value for human development, an increasingly large body of literature has been devoted to examining the complementary relationship between education on health. Does education affect health?

The empirical answer to this question has important implications. In theory, it can directly test the model of demand for health capital, which predicts that education improves health

(Grossman, 1972). In policy, if education does have a large, beneficial effect on health, then educational interventions might serve as a more cost-effective tool for promoting population health than merely increasing public health spending.

A large, positive correlation between education and health has been observed extensively in many countries, no matter how education and health are measured.1 However, there is a

broad disagreement on how large the effect is and whether this effect is causal. While some

studies report a significant education effect on health (e.g., Van Kipperslius, O’Donnell, & van

Doorslaer, 2011; Oreopoulos, 2006, 2007; Lleras-Muney, 2005), some other studies find small

or no effects (e.g., Meghir et al., 2018; Lager & Torssander, 2012; Clark & Royer ,2013;

Behrman et al.,2011; Arendt, 2005; Albouy & Lequien, 2009; Braakmann, 2011). Moreover,

some studies find mixed evidences across different health dimensions or sub-groups. Education

works for some health indicators or only for some specific groups (e.g., James, 2015; Kemptner

et al., 2011; Webbink et al., 2010).

Given the diversity of findings on the effects of education on health, the purpose of this

paper is to employ meta-analysis to quantitively review the empirical literature in this regard.

1 For a comprehensive review, see Grossman and Kaestner (1997), Grossman (2006), Cutler and Lleras-Muney (2008,2012). 2

Meta-analysis is a reliable and objective way to synthesize research findings and has been

extensively used in the field of economics, especially in the case that the empirical literature

lacks consensus (Neves et al., 2015; Stanley & Doucouliagos, 2012). It also employs

multivariate regression to reveal the factors underlying the heterogeneity of estimates, and to

establish whether there are any consistent and generalizable results which apply across contexts

and methods (Anderson et al., 2018).

To date, there are only two meta-analytic studies on the education effects on health.1

Furnee et al. (2008) is the first to conduct a meta-analysis on 88 estimates from 40 studies.

Their results show that the quality-adjusted life years (QALY) weight of a year of education is

0.036. However, their study only uses self-reported health as health measure. Since health is a multi-dimensional concept, self-reported health alone cannot delineate the overall picture of the health effects of education. Another study by Hamad et al. (2018) provide a review on the quasi-experimental studies of compulsory schooling laws. Their findings indicate that education has an effect on most health outcomes—most beneficial, some negative. However, they don’t make clear on what the effect size is. Since different studies use different health measures and estimation methodology, estimates may not be comparable in different studies.

Building on the previous studies, this paper seeks to answer three core questions. First, what is overall effect of education on health? Second, is there any publication bias in the current literature? Third, what are the factors explaining the heterogeneity of the estimates? This paper contributes to the existing meta-analytic literature in several ways. First, I carry out a systematic search and reporting procedure to extract standardized effect size estimates from different studies. The standardized effect size makes it possible to compare the different estimates in different studies. Second, the final sample consists of 4,671 estimates from 105 studies, representing the most comprehensive meta-analysis on the effects of education on

1 Another closely related literature is Galama et al. (2018). They reviewed the experimental and Quasi-experimental evidence on the effects of education on health and mortality. However, they didn’t conduct a formal meta-analysis. 3

health up to date. Third, I perform meta-regression analyses to reveal the factors behind the different estimated effects across studies. Fourth, as highlighted in some studies (e.g., Galama et al, 2018), heterogeneity may underlie the conflicting results on the effects of education on health. This study codes a detailed list of measures of health and education so as to explore the heterogeneous effects of education on health. Finally, I implement a variety of weighting procedures to calculate the mean effect, use clustered standard errors to eliminate serial correlations between estimates within studies, and conduct formal Funnel Asymmetry Tests

(FATs) and Precision Effect Estimate with Standard Error (PEESE) to detect publication bias.

I find that the overall effect of education on health is not economically meaningful, although statistically significant. There is substantial publication bias favoring a positive impact of education on health. The meta-regression analysis indicates that studies that do not address endogeneity are prone to exaggerate the effects of education on health. And the effect becomes weaker for more recent studies. The results cast doubt on the policy initiatives to improve population health through educational interventions.

The rest of the paper proceeds as follows. Section 2 introduces the conceptual framework.

Section 3 describes the meta-dataset. Section 4 conducts the preliminary analysis. Section 5 discusses publication bias. Section 6 explores the heterogeneity. Section 7 conducts robustness checks. Section 8 summarizes and concludes.

2. Conceptual Framework

In theory, there are several explanations why education may improve health. The first is the productive efficiency hypothesis, which states that people with higher education are more efficient producers of health because education increases the productive efficiency from the given quantities of health inputs (Grossman, 1972, 2006). The second is the allocative efficiency hypothesis, which argues that education can improve health through the optimal mix of health inputs. Better-educated people have more information on the deleterious effects of

smoking and bad habits, so that they are more likely to have healthy lifestyles (Rosenzweig &

Schulz, 1989). The third is that education improves health through channels such as better labor market opportunities, higher income, better living conditions, higher quality of care and living environment (Card, 1999; Cutler & Lleras-Muney, 2010).

In empirical literature, identifying the causal effects of education on health is plagued by endogeneity problems. The first originates from reverse causality. Healthier people usually have higher education (Behrman & Rosenzweig, 2004). Second, there may be omitted third variables including genetics, time preference and family background, which simultaneously cause both education and health to move in the same direction (Fuchs, 1982; Bijwaard et al.,

2015). For example, people with low time preference usually value future returns more highly, thus investing more in education and health concurrently.

Several identification strategies have been implemented to disentangle the causal relationships between education and health. However, the results of these studies are not consistent. The most widely used strategy is instrumental variable (IV). For example, Lleras-

Muney (2005), in her influential study, uses compulsory schooling laws as instruments for education and finds one additional year of compulsory schooling reduces 10-year mortality in the United States by as much as 6 percentage points. In contrast, Black et al. (2005), also using the compulsory schooling laws as instruments, find no effect of education on mortality among the US population. Furthermore, Buckles et al. (2016) exploits the exogeneous variation in years of college education caused by Vietnam draft and shows that education reduces mortality significantly. Finally, a recent study by Fletcher (2015) shows that the effects of education are not precisely estimated, though they appear to be large. It should also be noted that the variations in the results of instrumental variables studies may be partly due to the validity of instrument, as the exclusion condition of IV cannot be directly tested (Grossman, 2015).

The second approach to address endogeneity is Regression Discontinuity Design (RDD).

Compared to IV, RDD imposes weaker assumptions for identification. Oreopoulos (2006) employ compulsory schooling laws reform in UK as a discontinuity and finds a significant effect of education on both physical health and self-rated health. Clark and Royer (2013), however, use the same strategy and report no significant effect of education on mortality in

Britain. Albouy and Lequien (2009) also find that education does not have significant effect on the survival rate of French population. Meghir et al. (2018) examine the consequence of compulsory education reform in Sweden and find no significant effect of the reform on mortality.

The third alternative approach is twin fixed effects estimation. The logic behind this approach is that twins share almost the same characteristics such as genetic inheritance, gender, family characteristics and innate ability. It is plausible that differences in education between twins are exogenously determined, so differences in outcomes across twins can be seen as the outcomes from differences in their education. Using the Danish twin samples, Behrman et al.

(2011) find that education has no significant effect on mortality, while Van den Berg et al.

(2012) and Madsen et al. (2010) report mixed findings which depend on the sample studied.

On the basis of twin samples in Sweden, Lundborg et al. (2016) finds that education significantly reduces mortality. Furthermore, some studies show that education reduces the overweight for men (Webbink et al. 2010) but not for women (Webbink et al. 2010, Amin et al. 2013).

It can be seen from the prior literature that estimates on the effects of education on health vary greatly depending the methodology, health measures and sample used. In light of these conflicting findings, it becomes an empirical question to ascertain whether and to what extent education has an impact on health. Since there are large discrepancies in estimates across different studies, it is essential to synthesize these results and arrive at a general conclusion.

The meta-analysis can address this question.

3. The Meta-Data Set

3.1 Search Strategy and Selection Criteria

Following the MAER-Net guidelines proposed by Stanley et al. (2013), I carry out a systematic

search for the potential studies in the following online databases: EconLit, Web of Knowledge,

Google scholar, JSTOR, EBSCO, RePEc, IDEAS, SSRN, Scopus, ProQuest, NBER, IZA,

OECD Library and World Bank Publications. In line with Minasyan et al. (2019), reference

snowballing techniques was also employed to collect articles identified through the search

engine process. The search scope covers published journal articles, book chapters, conference

proceedings, working papers, master theses, doctoral dissertations, research reports, and other

technical reports. The combinations of the following key words are used: “education”,

“schooling”, “health”, “mortality”, “disease”, “obesity”, “BMI”, “morbidity”,

“depression”, “cognition”, “life expectancy” and “survival”. The search process was

completed at the end of December, 2018.

The initial search identified 474 studies in total. To make the analysis consistent, I adopt

a stepwise procedure to finalize the sample. The first step is to remove the duplicated studies,

including working papers reporting the same estimates with the final published version. Second,

I drop the studies which examined the role of health education, for example, the physical

education or oral health education, which was not directly related to our subject. Third, I delete

the studies that examine the effect of health on education. Forth, I restrict the sample to the

effect education on own health, excluding the studies investigating the intergenerational effect of education on health. Fifth, I discard studies lacking sufficient information to calculate t- statistics (that is, standard errors, t-statistics, p-value, or 95% confidence Intervals). Sixth, I exclude the theoretical studies or systematic reviews which did not report econometric estimates. Lastly, I eliminate studies which includes interaction terms and quadratic specifications of the education variable in the regression specification because it is difficult to

extract the partial estimated effect (Gunby, Jin & Reed, 2017). I collected 4,718 estimates from

107 studies. The detailed flowchart is shown in Appendix A. The studies are listed in Appendix

Table 1 reports the source of selected studies. The majority of the estimates (73.52%) is

extracted from journal articles, followed by estimates from working papers (24.62%),

conference proceedings (1.54%) and books (0.32%). With regard to the common journal outlets,

the journal with highest percentage of estimates is Social Science & Medicine (18.49%), an

interdisciplinary journal devoted to the exploration of social science on health. The next

journals in terms of frequency are Journal of Health Economics (10.83%), Economics &

Human Biology (9 %), Social Science Research (7.69%), and Economics of Education Review

(6.29%). Following these are a list of economic journals (OECD Journal: Economic Studies,

Health Economics, China Economic Review, Applied Economics), population journal

(Demography) and public health journal (International Journal of Epidemiology). This proves

that education-health nexus is a broad, appealing subject in the field of multi-disciplines.

3.2 Exclusion of Outliers

Outliers may be a concern in the meta-analysis. Before proceeding to code the studies, I

remove a few implausibly influential estimates from the dataset. Following Gallet &

Doucouliagos (2017), I first estimate a FAT-PET and then remove observations with absolute

value of the standardized residual greater than 3.5. As a consequence, the 47 outliers are

removed from the dataset. The final sample consists of 105 studies with 4671 estimates. Note

that the main results remain quantitatively similar without excluding the outliers.

Figure 1 displays the histogram of t-statistics used in the final sample. The distribution is

right-skewed with a few large outliers, indicating many studies reporting positive estimates. Of

these, 2183 (46.74%) recorded a significant, positive relationship between education and health

at the 10% significance level or below. 2319 (49.65%) record no significant relationship. There

are 169 estimates (3.62%) reporting a significant, negative relationship.

3.3 Coding Procedure

For each study in the sample, I code the data to extract a range of study characteristics.

This includes: the study’s author(s), publication year, publication status (e.g., journal, working paper, books), journal name, data type (cross-sectional or panel; individual level or aggregate level), countries studied, names of health variables, names of education variables, number of observations, and sample type (whole population or subsamples). I also record the regression coefficient and its associated standard errors or confidence intervals so as to derive the partial correlation coefficient (PCC).

It is worth noting that there are a number of studies not reporting the t-statistics or the standard error but p-values. I use the TINV function in Excel to calculate the t-statistics using the p-values and the degrees of freedom (Stanley & Doucouliagos, 2012). Similarly, there are some studies only reporting the levels of statistical significance with asterisk mark ***, **and

*. I follow the rule of thumb to assume that p-value is 0.01 for ***, 0.05 for **, 0.1 for * and

0.5 for no asterisk mark. I then use these p-values and degrees of freedom to calculate the t- statistics.

In addition, some studies employ non-linear estimations (e.g.,Probit/logit, cox proportional hazard model) and report only the Odds Ratio (OR) with standard error or 95%

confidence intervals. In these cases, I calculate t-statistics by 𝑡 or 𝑡 ..

,𝑤ℎ𝑒𝑟𝑒 𝑆.𝐸. . .. ∗.

The following ten types of estimation methods are coded: OLS, Feasible generalized least squares (FGLS), probit/logit, ordered logit or probit, Hierarchical Linear Model (HLM), instrumental variables (IV), Fixed Effects (FE), Regression Discontinuity Design (RDD),

Experiment/Quasi experiment, other estimation techniques. A set of dummy variables are also

generated to indicate whether some variables were controlled in the regression analysis, which

include age, gender, race, marital status, income and occupation.

As previous mentioned, endogeneity is one of the uttermost concerns in the education and

health literature. For this reason, I pay particular attention to estimates produced by estimation

methodology addressing endogeneity (IV, FE, RDD, Experiment/Quasi experiment). I will

examine whether endogeneity matters a lot for the effect of education on health.

I also code the methods used to derive t-values into a set of dummies (tNormal,

tCalculatedBypValue and tCalculatedByCI). The type of standard error associated with estimates are coded as SEspherical and SEnonspherical.

TABLE 3 presents the detailed information on the health measures in the sample. I divide

the health measures into three broad categories: physical health, mental health and general

health. Of physical health, the most frequent measure is mortality (25.37%). The second

frequent measure of physical health is obesity (21.43%), which includes BMI, overweight,

body size etc. The third frequent measure of physical health is the presence of a particular

illness or disease such as hypertension, heart disease, diabetes (19.72%). The last measure of

physical health is Activities of Daily Living (8.07%).

Of measures of mental health, depression is the common form of mental health problems

(6.83%). Another type of mental health is cognition (0.77%). Many studies used instruments

such as CES-D and malaise score to measure mental well-being. The majority of mental health

measures is self-reported.

For measures of general health, it is common to elicit individual’s general health through

five-point scale subjective assessment: very poor, poor, fair, good and excellent. The general

health is generally assessed through self-reporting.

TABLE 4 displays the education variables in the sample. 45.94% of education variable is

coded as continuous years of schooling, and 54.06% of education variable is coded as

categorical educational levels. Of categorical education levels, about a half (49.94%) is

secondary education. Primary and tertiary educations account for 19.52% and 30.53%

respectively.

4. Preliminary Analysis

4.1 Effect Sizes

The first step in meta-analysis is to extract the estimates from the literature. To do that, I

focus on studies that estimate the effects of education on health with the following regression

specification:

H𝛽 𝛽𝐸𝑑𝑢𝑐 ∑ 𝛽𝑋 error (1)

Where H is a measure of health, 𝐸𝑑𝑢𝑐 is a measure of education, 𝑋is a vector of

control variables, and 𝛽 is the parameter to be estimated. In meta-analysis, 𝛽 is the

dependent variable, representing the estimated effect of education on health in study 𝑖.

However, 𝛽 cannot be directly used in meta-analysis because the estimates are not

comparable across different studies. This is due to the differences in the measure of education and health, or estimation methods. For example, some studies used continuous years of schooling as a proxy for education, while others used categorical education levels. The health measures also differ across studies. Furthermore, because different estimation procedures are employed in different studies, the coefficients were interpreted in various ways. To circumvent this problem, I follow the common approach and use partial correlation coefficient (PCC) to

convert estimates into a unitless, comparable measure. PCC is specified as follows:

𝑡 𝑃𝐶𝐶 𝑡 𝑑𝑓

Where 𝑡 is t-statistics, 𝑑𝑓 is degree of freedom. The value of PCC lies between -1 and

The corresponding standard error of 𝑃𝐶𝐶 is calculated by:

1𝑃𝐶𝐶 𝑠. 𝑒. 𝑃𝐶𝐶 𝑑𝑓

Accordingly, 𝑃𝐶𝐶 and 𝑠. 𝑒. 𝑃𝐶𝐶 will be used in the following meta-analysis.

Figure 2 depicts the distribution of PPC values in the sample. Compared to the distribution

of t-statistics, the distribution of PCC displays a normal shape. The mean and median value in our sample are 0.027 and 0.015. According to the guidelines proposed by Doucouliagos (2011),

PCC less than 0.07 can be considered to be small, with 0.17 or above considered to be moderate, and 0.33 or above large. Judged by this criterion, the effects of education on health is very small. Nevertheless, this is not surprising because almost half of the estimates (49.65%) in the sample is statistically insignificant.

4.2 Calculation of Mean effects

To compute the mean effect of education on health, I employ two commonly used estimators in meta-analysis: the fixed effects (FE) estimator and random effects (RE) estimator. 1

The FE implicitly assumes that there is a single underlying true effect and the reason for

different estimates across studies is because of sampling error. The FE estimator can be

estimated by Weighted Least Square (WLS), with weights being the inverse of 𝑠. 𝑒. 𝑃𝐶𝐶. It

can be formulated as follows:

, , 𝑖 1,2, … ,𝑁 (2) 𝑠.𝑒.𝑃𝐶𝐶𝑖 𝑠.𝑒.𝑃𝐶𝐶𝑖 𝑠.𝑒.𝑃𝐶𝐶𝑖

Where 𝑁 is the number of estimates in the dataset. 𝛼 measures the mean effect of

education on health.

Some scholars argued that the assumption underlying FE is too restrictive. There may be not a single underlying effect, as each study may has its own true effect size. In meta-analysis,

1 It should be noted that the terms “Fixed Effects” and “Random Effects” are not the same concepts as in the panel data in econometrics. 12

this heterogeneity can be accommodated by RE estimator.

In the RE estimator, the variation in the effect sizes consists of two parts: the sampling

error and the heterogeneity in the true effect size. Assume the heterogeneity represented by 𝜏 independent of sampling error, the variation can be expressed as:

𝜔 𝑠. 𝑒. 𝑃𝐶𝐶 𝜏 (3)

In this case, RE estimator is weighted average of all the estimates of the effect size in

studies, with the weight given by 𝜔. It can be obtained by:

, ,𝑖1,2,…,𝑁 4

Which method should be preferred in meta-analysis? There is no consensus among

researchers yet (Reed, 2015; Doucouliagos & Paldam, 2013; Stanley & Doucouliagos, 2012).

Although it is generally held that the assumption of RE model is more realistic, some scholars

argue that the FE estimator produce less biased estimates if there is publication bias. As a

consequence, both estimates will be reported in this study.

Another issue relates to the weight schemes in this study. In the dataset, the number of

estimates per study varies greatly, ranging from 1 to 238, with a mean of 44. The preceding

analysis implicitly gives more weight to studies with more estimates and less weight to studies

with few estimates. In line with previous studies (Xue & Reed, 2019; Gunby, Jin & Reed,

2017), I employ two weighting schemes, i.e., equal weight to studies (weight 1) and equal

weight to individual estimates (weight 2) in the following analysis.

Lastly, since there are more than one estimates in most studies in the sample, it is likely

that the estimates are correlated within the study. This will lead to serial correlation and

inefficient estimator. To mitigate this concern, I will estimate the model by Weighted Least

Square (WLS) with heteroscedasticity cluster-robust standard errors, which accommodates the

serial correlation between estimates within each study.

Table 5 reports the preliminary results of the mean effect of education on health. As is

shown, the estimated effects of education on health lie between 0.012 to 0.03 and are significant

at 5% level. Taking all estimates into account, the overall PCC between education and health

is approximately 0.02. According to the guidelines for assessing the strength of a correlation

coefficient proposed by Dougcouliagos (2011), the mean effect is far below 0.07, thus being

very small, although statistically significant. It should be noted that the simple overall meta-

analysis should be interpreted with caution in the case of publication bias. In next section, I

will investigate whether there is publication bias and how it might affect the reported estimates

in the current literature.

5. Publication Bias

In meta-analysis, publication bias poses a serious threat to the validity of analytic results.

Publication bias refers to the fact that peer-reviewed journals are more likely to publish studies

with significant results than studies with nonsignificant results, or some studies with

insignificant results are seldom written out by authors. Publication bias will result in the

incorrect estimates of the mean effect sizes.

An informal way to detect publication bias is the funnel plot, which plots the effect size

on the x-axis and standard error on the y-axis (Egger et al., 1997). If there is no publication

bias, the distribution of the standard error will be symmetric around the mean line. Publication bias introduces asymmetry into the funnel plot. In the presence of an upward bias, the scatter- dot will cluster on the right of the mean line, or vice versa.

Figure 4 shows that funnel plot of the PCC and standard error in the sample. It can be seen that as the standard error increases, the PCC values are skewed to the right, suggesting an upward publication bias towards a positive impact of education on health.

In addition to the visual representation of the publication bias, a formal way to test publication bias is Funnel-Asymmetry-Precision Test (FAT-PET) (Stanley, 2005; Stanley,

2008). FAT-PET is a simple meta-regression of the PCC on its standard error:

𝑃𝐶𝐶 𝛼 𝛼𝑠𝑒𝑃𝐶𝐶𝜐 (5)

Equation (5) can be estimated by FE estimator and RE estimator, as in Equation (2) and

Equation (4).

Based on Monte Carlo simulations, Stanley (2008) further argues that 𝛼 in Equation (5) may be biased downward in the case that null hypothesis is rejected. To surmount this problem,

I follow the suggestion of Stanley and Doucouliagos (2014) and replace the standard error with its square term. In this case, 𝛼 is called the Precision Effect Estimate with Standard Error

(PEESE) which is specified as:

𝑃𝐶𝐶 𝛼 𝛼𝑠𝑒𝑃𝐶𝐶 𝜐 (6)

Equivalently, Equation (6) can also be estimated by FE estimator and RE estimator as in

Equation (2) and Equation (4).

Table 6 reports the results of FAT-PET. All the four columns reject the null hypothesis

𝐻0: 𝛼1 = 0 at the 5 % significance level, suggesting the presence of publication bias. The positive FAT coefficients suggest upward publication bias, indicating that the current literature favors the publication of positive impacts of education on health. In three of the four columns,

I also reject the null hypothesis: 𝛼0 = 0, with the estimates of 𝛼0 at least significant at the 10% level. The only exception is the “Fixed Effects (Weight2)” regression. In general, the PET coefficients suggest that education is positively correlated with health. However, bias-adjusted estimates of the mean true effect of education on health range from 0.008 to 0.015, far below the value that Doucouliagos (2011) identifies as being “small”. Therefore, the overall effect of education on health is not economically meaningful.

Table 7 further presents the PEESE results. Compared to Table 6, all the four columns reject the null hypothesis 𝛼1 = 0, 𝛼0 = 0. The effects of bias become larger and significant at

1% level, further confirming the persistent publication bias in the literature. The precision effect become more significant and slightly larger, ranging from 0.011 to 0.024. Again, these

values are still far below the threshold that Doucouliagos (2011) identifies as being “small”. In

sum, Table 6 and Table 7 consistently support the point that the effect of education on health

is very small and the current literature suffers from substantial publication bias.

6. Modelling Heterogeneity

In this section, I employ meta-regression analysis to explore why the estimates vary

systematically across different studies. The differences in the reported estimates may stem from

model specification, research design or from differences across countries and over time. To do that, I estimate the following regression s:

𝑃𝐶𝐶 𝛼 𝛼𝑠𝑒𝑃𝐶𝐶∑ 𝛼 𝑋 𝜀 (7)

Where 𝑋 is the vector of moderator variables, 𝛼 is the coefficients, 𝑠𝑒𝑃𝐶𝐶 is the standard error of 𝑃𝐶𝐶 , 𝜀 is the error term. Vector 𝑋 contains the following variables:

Measures of Health: Seven measures of health are included in the regression analysis.

They are: general health, ADL, disease, mortality, obesity, mental health and whether the

health is self-reported. Dummy variables are created to represent each measure of health.

Mortality, Obesity and Disease are three most popular measures of health in the dataset, with

25.37%, 21.43% and 19.72% respectively. General health is also a popular measure,

accounting for 17.81% of the estimates. However, mental health is less commonly used, with

only 7.6% in the dataset. The majority of health measures is self-reported (63%).

Measures of Education: Both the continuous and categorical education measures are used

in the literature. The continuous education measure is years of schooling. Categorical education

measures are coded into three levels: primary, secondary and tertiary. 46% of the estimates use

continuous years of schooling. The remaining estimates (54%) are categorical education levels.

The mean number of education variables in the dataset is 1.866.

Regions: The main countries/regions in the dataset cover North America, Europe, East

Asia Pacific and some other countries. 35.4% of the estimates are based on North American

and 51.5 % from Europe. Asian countries account for 12.1% of the estimates.

Income level: A country’s income level is coded on the basis of UN classifications1. Most

estimates (91.5%) use data from high-income counties. Only a small fraction (8.5%) use data

from middle-low income countries.

Data characteristics: most studies use individual-level data (97.8%) and panel data

(72.2%). I will explore whether data characteristics influence the reported results.

Sample characteristics: There are five categories of samples: whole population, male sample, female sample, sample aged 25 to 50, sample aged 50 or above, and sample mixed with gender and age. About half of the estimates (48.8%) is based on the whole population. I will investigate whether the effects differ across different samples.

Control variables: Most studies control a set of variables. I include six common variables:

Age, Gender, Race, Marital Status, Income and Occupation. The variables are coded as 1 if they are included in the regression equation as explanatory variables and 0 if otherwise. Most studies control gender and age in regression specification.

Endogeneity: While many studies use OLS or non-linear models, some studies control for the endogeneity by employing Instrumental Variable (IV), Regression Discontinuity (RD),

Fixed Effects and Experiment/Quasi-experiment methods. To be concise, I code the studies controlling for endogeneity as 1 and 0 if otherwise. 30.2 % of the estimates are based on the studies addressing endogeneity. I will investigate whether the effect sizes differ once the endogeneity is controlled for.

Calculation of T-statistics and Standard Error. I also code the methods to calculate the t- statistics. There are three set of dummy variables: tNormal, tCalculatedBypValue, tCalculatedByCI. Most of the T-statistics is calculated by normal ways (68%), followed by

confidence interval (19%) and P-value (13%). Moreover, 49.2% of the estimates assume non-

1 The details can be accessed at: https://www.un.org/en/development/desa/policy/wesp/wesp_current/2014wesp_country_classification.pdf 17

spherical standard errors.

Publications type: The final set of variables relate to different dimensions of publication

process. A dummy variable Unpublished is created to account for the difference between

published and unpublished studies. Four sets of dummy variables (Economics Journal,

Sociology Journal, Population & development journal, Public health & Medicine Journal,

Science Journal) are created to indicate the different journal types I also explore the effect of

quality of journal through JournalRank, which is based on the Scopus Citescore Metrics. The

continuous Pubyear is set to quantify the time trend of the effect sizes.

Table 8 lists all explanatory variables, and their mean values, minimum and maximum.

Equation (7) is estimated by FE and RE with weight 1 and weight 2.

Given the large number of variables included in the dataset, multicollinearity may arise

and confound significant relationships. Moreover, there is model uncertainty on which

variables should be included in the regression specification. To overcome these problems, I

choose to use variable selection procedure developed by Lindsey & Sheather (2010) to select

the best set of control variables. 1The advantage of this procedure is that it avoids the arbitrary

selection of variables. On the basis of the information criteria such as Akaike’s Information

Criterion (AIC), Akaike’s corrected Information Criterion (AIC) and Bayesian Information

Criterion (BIC), it helps researchers obtain the optimal model by dropping redundant variables

iteratively (Amin, 2016). This selection procedure is performed by invoking the command

“vselect” in Stata. I employ the backward elimination procedure to perform the selection

process to determine the best model specification with the smallest BIC value. I also lock the

publication bias variable (SE) and a variable that represent attempts to address endogeneity

(endogeneity) in each round of the selection process.

1 Another popular approach for model selection is Bayesian Model Averaging (BMA), which has been widely used in the meta-analysis literature. However, BMA is very susceptible to weighting schemes (Gunby, Jin & Reed, 2017). This paper uses variable selection procedure instead. 18

In the initial round, all the 39 moderator variables (excluding reference variables) are

included in the regression specification. At each subsequent round, the variable selection

procedure eliminates the variable that results in the largest reduction in BIC. It keeps on doing

that until the BIC can no longer be minimized. After obtaining the best set of moderator

variables, I re-estimate the model by FE and RE estimators with weight 1 and weight 2.

The main findings are presented in Table 9. The SE term is positive and statistically

significant at least at the 5 percent level in four estimations. The positive, significant

coefficients of SE after controlling for a set of moderator variables suggests that the FAT results from TABLE 7 are not a spurious outcome caused by omitted variables. The education-health literature suffers from substantial publication bias.

Not all the health variables are consistently significant across all four estimations. The coefficients on disease are negative and always significant, indicating the effects of education on disease are weak compared to the effects of education on general health. This finding underlines the heterogeneity in the effects of education on health. Moreover, the coefficients on ADL, mortality and depression are sensitive to weight option and not consistently significant, implying no significant difference between the effects on general health, ADL, mortality and depression. The positive coefficients of SelfReported in three of the four columns suggest that studies with self-reported health measures tend to overstate the impact of education on health.

With regard to education measures, compared to primary education, continuous years of schooling appears to have a larger and significant effect on health. In contrast, secondary education and tertiary education do not seem to make a clear difference compared with primary education.

Whether the studies correct for endogeneity or not systematically affects estimates.

Holding other variables constant, controlling for endogeneity will reduce the PCC values by

0.026~0.037, which are all significant at 1% level. The further analysis of the sample that

addresses endogeneity shows that none of the mean effects in FAT-PET-PEESE estimations

exceeds 0.01.1

The last variable that is consistently significant across all four estimations is pubYear. The negative coefficients suggest that the estimated effects of education on health decrease with time.

The results for the other moderator variables in Table 9 are more inconclusive. In the following discussion, I focus on those variables that are significant at 10% level in three of the four regressions. The coefficients for sample_male and sample_other are negative and significant, suggesting that men and other samples reap less health returns from education compared to the whole population. This finding is not surprising because men are more likely to have unhealthy behaviors, which may reduce the health benefits of education.

The negative coefficient of tCalculatedbyCI indicates that the estimated effect of education on health tend to be smaller if the t-statistics is calculated by confidence intervals.

One possible concern is that the mean effect may be downward biased by including estimated effects using these t-statistics. Table 9 shows that the coefficient ranges from -0.001 to -0.044.

However, after accounting for this effect, the mean effect size is still below the threshold 0.07 value which Doucouliagos (2011) identifies as “small”.

Regarding publication characteristics, compared to non-journal publications, economics journal and sociology journal seem to publish more negative findings. The positive coefficient estimates of Journalrank suggest that high-quality journals report more positive effects of education on health. Neither of other moderator variables show consistent significance across the four estimations.

1 However, Galama et al. (2018) and Grossman (2015) argue that the smaller effect after the endogeneity is addressed may be due to the fact that IV estimation produces local average treatment effect (LATE) rather than average treatment effect (ATE). That is, the affected population is different. 20

7. Robustness Checks

I carry out robustness checks on previous findings in two ways. First, following Stanley &

Doucouliagos (2012), I use an alternative effect size calculation and test for the robustness of the main results by implementing z transformations of the PCCs and the standard errors. The results of the FAT-PET-PEESE based on Z transformation are shown in TABLE 10. It can be seen that, compared to TABLE 6 and TABLE 7, the main results in TABLE 10 remain unchanged. Second, I re-estimate the FAT-PET-PEESE using the full sample. The results in

TABLE 11 remain quantitatively similar to the results with truncated sample. Therefore, across all the robustness checks, the main qualitative conclusions from above are still valid.

8. Conclusion

The education-health nexus has been an intriguing topic in the field of economics and other disciplines for many years. A large body of literature has been emerging over the past decades to examine the education effects on health. However, no consensus has been reached on how large the effect is and whether the effect is causal.

This paper provides a meta-analysis on the extensive literature that examine the impact of education on health. The final sample consists of 4,671 estimates from 105 studies. The main findings indicate that the overall effect of education on health is very small, although statistically significant. This finding echoes Clark & Royer (2013) and Meghir et al., (2018) in that education plays no big role in the process of health production. I also find that there is severe publication bias favoring a positive impact of education on health. The meta-regression analysis indicates that studies that do not control for endogeneity are prone to exaggerate the effects of education on health. The effects become smaller for more recent studies. Thus, this study implies that the theory of demand for health capital that assume a positive role of education in health deserves further investigation.

In terms of policy implications, education has been proposed as one of important health

policy initiatives in countries including US and UK (Clark & Royer, 2013). The similar

initiative also appears regularly in international organizations such as OECD and WHO (OECD,

2010; WHO, 2015). However, the findings of this paper cast doubt on the feasibility of policies

designed to improve health through educational interventions.

To sum up, this study represents the first comprehensive meta-analysis to identify the

overall effect of education on health. However, there are still many unknowns regarding the

complex relationship between education and health. One important area for future work is to

better understand the mechanisms underlying the discrepancies among the estimates across

different studies. Another fruitful direction is to examine the effects of education quality on

health. The last direction, as cautioned by Grossman (2015), is that the validity of instruments in the current literature needs to be assessed. It is anticipated that this paper will stimulate more research in the future.

References

Albouy, V., Lequien, L. (2009). Does compulsory education lower mortality? Journal of

Health Economics, 28, 155–68.

Amin, A. (2016). Exploring the role of economic incentives and spillover effects in biodiversity

conservation policies in sub-Saharan Africa. Ecological Economics, 127, 185-191.

Amin, V., Behrman, J.R., Spector, T.D. (2013). Does more schooling improve health outcomes

and health related behaviors? Evidence from U.K. twins. Economics of Education

Review ,35,134-148.

Anderson, E., d’Drey, M.A.J., Duvendack, M., Esposito, L. (2018). Does government spending

affect income poverty? A Meta-regression Analysis. World Development. 103, 60-71.

Arendt, J. N. (2005) Does education cause better health? A panel data analysis using school

reforms for identification. Economics of Education Review, 24, 149–60.

Becker, G. S. (1964). Human capital. New York, USA: NBER.

Behrman, J.R., Rosenzweig, M.R. (2004). Returns to birthweight. Review of Economics and

Statistics, 86 (2), 586–601.

Bijwaard, G.E., Kippersluis, H., Veenman, J. (2015). Education and Health: The role of

cognitive ability. Journal of Health Economics, 42:29-43.

Black, D., Hsu, Y.C., Taylor, L.J. (2015). The effect of early-life education on later-life

mortality. Journal of Health Economics,44,1-9.

Braakmann, N. (2011). The causal relationship between education, health and health related

behavior: evidence from a natural experiment in England. Journal of Health Economics,

30, 753–63.

Buckles, K, Hagemann, A., Malamud, O., Morrill, M., Wozniak, A. (2016). The effect of

college education on mortality. Journal of Health Economics, 50,99-114.

Card, David E. 1999. The Causal Effect of Education on Earnings. In Handbook of Labor

Economics,Vol. 3A, edited by Orley Ashenfelter and David Card, 1801-1863. Amsterdam:

Elsevier.

Clark, D., & Royer, H. (2013). The effects of education on adult mortality and health: evidence

from Britain. American Economic Review, 103 (6), 2087–2120.

Cutler, D. M., & Lleras-Muney, A. (2006). Education and Health: Evaluating Theories and

Evidence. NBER Working Paper 12352.

Doucouliagos, H. (2011) How large is large? Preliminary and relative guidelines for

interpreting partial correlations in economics. Economics Series 2011/5, Deakin

University, Faculty of Business and Law, School of Accounting, Economics and Finance.

Doucouliagos, H., Paldam, M. (2013) The robust result in meta-analysis of aid effectiveness:

A response to Mekasha and Tarp. The Journal of Development Studies, 49(4): 584-587.

Egger, M., Davey Smith, G., Schneider, M. and Minder, C. (1997). Bias in meta-analysis

detected by a simple, graphical test. British Medical Journal 315: 629–634.

Fletcher, J. M. (2015). New evidence of the effects of education on health in the US:

Compulsory schooling laws revisited. Social Science & Medicine, 127, 101-107.

Fuchs, V.R. (1982). Time preference and health: an exploratory study. In: Fuchs, V. (Ed.),

Economic Aspects of Health. The University of Chicago Press, Chicago.

Furnee, C. A, Groot, w., van den Brink, H. (2008). The health effects of education: a meta-

analysis. European Journal of Public Health. 18(4), 417–421.

Galama, T. J., Lleras-Muney, A., van Kippersluis, H. (2018). The effects of education on

Health and Mortality: A Review of Experimental and Quasi-Experimental Evidence.

NBER Working Paper No. 24225

Gallet, C.A., Doucouliagos, H. (2017). The impact of healthcare spending on health outcomes:

A meta-regression analysis. Social Science & Medicine, 179, 9-17.

Grossman, M. (1972). On the concept of health capital and the demand for health. Journal of

Political Economy ,80, 223–255.

Grossman, M., Kaestner, R. (1997). Effects of Education on Health, in J. R. Behrman and N.

Stacey (eds.) The Social Benefits of Education (Ann Arbor: University of Michigan

Press).

Grossman, M. (2006). Education and Nonmarket Outcomes. In Handbook of the Economics

of Education, Vol. 1, edited by Eric Hanushek and Finis Welch, 577–633. Amsterdam:

North-Holland.

Grossman, M. (2015). The relationship between health and schooling: what is new? Nordic

Journal of Health Economics, 3(1), 7-17.

Groot, W., Van den Brink, H. (2007). The health effects of education. Economics of Education

Review, 26, 186–200.

Gunby, P., Jin, Y. and Reed, W.R. (2017). Did FDI really cause Chinese economic growth? A

meta-analysis. World Development, 90, 242-255.

Hamad,R., Elserb,H., Tranc, D.C., Rehkopfc, D.H., Goodman, S.N. (2018).How and why

studies disagree about the effects of education on health: A systematic review and meta-

analysis of studies of compulsory schooling laws. Social Science & Medicine, 212,168-

178.

James, J. (2015). Health and education expansion. Economics of Education Review, 49, 193-

215.

Kemptner, D., Jürges, H., Reinhold, S. (2011). Changes in compulsory schooling and the causal

effect of education on health: evidence from Germany. Journal of Health Economics,

30,340–54.

Lager, A. C. J., & Torssander, J. (2012). Causal effect of education on mortality in a quasi-

experiment on 1.2 million Swedes. Proceedings of the National Academy of Sciences, 109

(22), 8461-8466.

Lindsey, C., Sheather, S. (2010). Variable selection in linear regression. Stata Journal. 10, 650-

669.

Lleras-Muney, A. (2005). The Relationship Between Education and Adult Mortality in the

United States. Review of Economic Studies,72 (1): 189-221.

Meghir, C., Palme, M., & Simeonova, M. (2018). Education and Mortality: Evidence from a

Social Experiment. American Economic Journal: Applied Economics, 10(2): 234-256.

Minasyan, A., Zenker, J., Klasen, S., Vollmer, S. (2019). Educational gender gaps and

economic growth: A systematic review and meta-regression analysis. World Development.

122,199-217.

Neves, P.C., Afonso, O., Silva, S.T. (2016). A Meta-analytic Reassessment of the Effects of

Inequality on Growth. World Development. 78, 386-400.

OECD. (2010). Improve health and social cohesion through education. Paris, OECD.

Oreopoulos, P. (2006) Estimating average and local average treatment effects of education

when compulsory schooling laws really matter. American Economic Review, 96, 152–75.

Oreopoulos, P. (2007) Do dropouts drop out too soon? Wealth, health and happiness from

compulsory schooling. Journal of Public Economics, 91, 2213–29.

Reed, W.R. (2015). A Monte Carlo analysis of alternative meta-analysis estimators in the

presence of publication bias. Economics: The Open-Access, Open-Assessment E-

Journal,9 (2015-30): 1-40.

Rosenzweig, M.R., Schultz, T.P. (1989). Schooling, information and nonmarket productivity:

contraceptive use and its effectiveness. International Economic Review, 30, 457–477.

Schultz, T. W. (1961). Investment in human capital. American Economic Review, 51(1), 1-17.

Stanley, T. D. (2005). Beyond publication bias. Journal of Economic Surveys, 19(3),309–345.

Stanley, T. D. (2008). Meta-regression methods for detecting and estimating empirical effect

in the presence of publication selection. Oxford Bulletin of Economics and Statistics, 70(1),

103–127.

Stanley, T.D., Doucouliagos, H. (2012). Meta-regression analysis in economics and business.

Abingdon: Routledge.

Stanley, T. D., Doucouliagos, H., Giles, M., Heckemeyer, J. H., Johnston, R. J., Laroche, P.,

Rost, K. (2013). Meta-analysis of economics research reporting guidelines. Journal of

Economic Surveys, 27(2), 390–394.

Stanley, T.D., Doucouliagos, H. (2014) Meta-regression approximations to reduce publication

selection bias. Research Synthesis Methods, 5, 60–78.

United Nations. (2015). Transforming our world: the 2030 Agenda for Sustainable

Development.

Van Kippersluis, H., O’Donnell, O. & Van Doorslaer, E. (2011). Long-run returns to education:

Does schooling lead to an extended old age? Journal of Human Resources, 46 4 695-721.

Webbink, D., Martin, N. G. and Visscher, P. M. (2010). Does education reduce the probability

of being overweight? Journal of Health Economics, 29, 29–38.

WHO. (2015). Health 2020: Education and health through the life-course. Geneva, WHO.

Xue, X.D., Reed, W. R. (2019). Social Capital and Health: A Meta-Analysis.Working Papers

in Economics 19/01, University of Canterbury, Department of Economics and Finance.

TABLE 1 Source of Selected Studies

Source Estimates Percentage

Journal Articles 3434 73.52

Working Papers 1150 24.62 Books 15 0.32

Conference Papers 72 1.54 Total 4,671 100

NOTE: Author’s calculations.

TABLE 2 Common Journal Outlets

Journal Estimates Percentage

Economics Journal

Journal of Health Economics 372 10.83 Economics & Human Biology 309 9 Economics of Education Review 216 6.29 OECD Journal: Economic Studies 170 4.95 Health Economics 161 4.69 China Economic Review 129 3.76 Economics Perspectives 116 3.38 Journal of Population Economics 103 3 Applied Economics 84 2.45 Sociology Journal Social Science & Medicine 635 18.49 Social Science Research 264 7.69 Population Journal Demography 154 4.48 Public Health Journal American Journal of Epidemiology 93 2.71 International Journal of Epidemiology 83 2.42 NOTE: Authors’ calculations.

TABLE 3 Health Measurements

Measure Estimates Percentage

Physical Health 3484 Mortality 1185 25.37 Obesity 1001 21.43 Disease 921 19.72 ADL/IADL 377 8.07 Mental Health 355 Depression 319 6.83 Cognition 36 0.77 General Health 832 17.81 Total 4,671

NOTE: Authors’ calculations.

TABLE 4 Categorization of Education

Measure Estimates Percentage

Continuous (years of schooling) 2146 45.94 Categorical (education levels) 2525 54.06 Primary education 493 10.6 Secondary education 1261 27

Tertiary education 771 17

NOTE: Author’s calculations.

TABLE 5 Mean Effects without Correcting for Publication Bias Fixed Effects Fixed Effects Random Effects Random Effects Statistics (Weight1) (Weight2) (Weight1) (Weight2) 0.012*** 0.014** 0.025*** 0.03*** Mean (3.43) (2.33) (9.73) (10.98) Observations 4,671 4,671 4,671 4,671 NOTE: Columns (1)-(4) report the mean value for PCC. T-statistics is reported in parenthesis. All of the results are estimated by Weighted Least Squares (WLS) with cluster robust standard errors. * p<0.1, ** p<0.05, *** p<0.01.

TABLE 6 The Funnel Asymmetry Test (FAT) and Precision Effect Test (PET) Fixed Effects Fixed Effects Random Effects Random Effects Statistics (Weight1) (Weight2) (Weight1) (Weight2) 0.008* 0.011 0.012*** 0.015*** PET (𝛼 ) (1.82) (1.6) (4.02) (4.07) 1.6*** 1.943** 1.131*** 1.403*** FAT (𝛼 ) (2.75) (2.31) (4.28) (4.5) Observations 4,671 4,671 4,671 4,671 NOTE: All the results are estimated by Weighted Least Squares (WLS) with cluster robust standard errors as described in equation (5). T-statistics is reported in parenthesis. * p<0.1, ** p<0.05, *** p<0.01.

TABLE 7 Precision Effect Estimate with Standard Error (PEESE) Fixed Effects Fixed Effects Random Effects Random Effects Statistics (Weight1) (Weight2) (Weight1) (Weight2) 0.011*** 0.013** 0.021*** 0.024*** Precision (𝛼 ) (3.16) (2.25) (8.01) (8.77) 48.29*** 52.18*** 19.6*** 26.08*** Bias (𝛼 ) (3.51) (3.43) (2.77) (3.8) Observations 4,671 4,671 4,671 4,671 NOTE: All the results are estimated by Weighted Least Squares (WLS) with cluster robust standard errors as described in equation (6). T-statistics is reported in parenthesis. * p<0.1, ** p<0.05, *** p<0.01.

TABLE 8 Description of Variables

Variable Description MeanMin Max

EFFECT SIZE PCC Partial Correlation Coefficient 0.027 -0.145 0.275 S.E. The standard error of the PCC 0.014 0.001 0.123 HEALT MEASURES GeneralHealth =1, if health measure is general health (Reference) 0.178 0 1 ADL =1, if health measure is activities of daily living or health limitations 0.081 0 1 Disease =1, if health measure is specific diseases 0.198 0 1 Mortality =1, if health measure is mortality 0.254 0 1 Obesity =1, if health measure is obesity 0.214 0 1 MentalHealth =1, if health measure is mental health 0.076 0 1 SelfReported =1, if self-reported health 0.63 0 1 EDUCATION MEASURES Primary =1, if education measure is primary education (Reference) 0.106 0 1 Secondary =1, if education measure is secondary education 0.27 0 1 Tertiary =1, if education measure is tertiary education 0.165 0 1 Education attainment =1, if education measure is years of education 0.46 0 1 NumberofEducvariables Number of Education variables 1.866 1 8 COUNTRIES NorthAmerica =1, if countries in North America included (Reference) 0.354 0 1 Europe =1, if countries in North Europe included 0.515 0 1 AsiaPacific =1, if countries in Asia and Pacific included 0.121 0 1 OtherCountry =1, if countries studied is OCED or transnational 0.01 0 1

Variable Description MeanMin Max

INCOME LEVEL Middle-Low_Income =1, if country studied is a middle or low-income country (Reference) 0.085 0 1 HighIncome =1, if country studied is a high-income country 0.915 0 1 DATA CHARACTERISTICS Aggregate =1, if estimate is from aggregate-level data (Reference) 0.022 0 1 Individual =1, if estimate is from individual-level data 0.978 0 1 Cross-sectional =1, if estimate is from cross-sectional data (Reference) 0.278 0 1 Panel =1, if estimate is from panel data 0.722 0 1 SAMPLE CHARACTERISTICS WholeSample =1, if estimate is from whole population (Reference) 0.488 0 1 FemaleSample =1, if estimate is from female population 0.22 0 1 MaleSample =1, if estimate is from male population 0.241 0 1 Age25to50Sample =1, if estimate is from populations aged 25 to 50 0.022 0 1 Age50aboveSample =1, if estimate is from populations aged 50 and above 0.022 0 1 OtherSample =1, if estimate is from none of the above 0.007 0 1 CONTROL VARIABLES Age =1, if age is a control variable in the regression 0.681 0 1 Gender =1, if gender is a control variable in the regression 0.806 0 1 Race =1, if race is a control variable in the regression 0.286 0 1 MaritalStatus =1, if marital status is a control variable in the regression 0.228 0 1 Income =1, if income is a control variable in the regression 0.133 0 1 Occupation =1, if occupation is a control variable in the regression 0.1 0 1 ESTIMATION METHOD Endogeneity =1, if endogeneity is addressed, including Instrumental Variable, Regression 0.302 0 1

Variable Description MeanMin Max

Discontinuity Design, Fixed Effects, Experiment/Quasi-Experiment. CALCULATION OF T STATISTIC tNormal =1, if t-statistic=coefficient/standard error (Reference) 0.68 0 1 tCalculatedBypValue =1, if t-statistic is derived from p-values 0.13 0 1 tCalculatedByCI =1, if t-statistic is derived from confidence intervals 0.19 0 1 SENonspherical =1, if standard error is non-spherical 0.492 0 1 PUBLICATION TYPE Unpublished =1, if unpublished (Reference) 0.265 0 1 Economics Journal =1, if published in economics journal 0.41 0 1 Sociology Journal =1, if published in sociology journal 0.195 0 1 Development & Population Journal =1, if published in development and population journal 0.06 0 1 Public health & Medicine Journal =1, if published in public health and medicine journal 0.057 0 1 Science Journal =1, if published in science journal 0.013 0 1 JournalRank Scopus Citescore Metrics, 0 for non-journal publications 2.136 0 8.58 PubYear Year study was published/appeared 2011.5 1987 2018

TABLE 9 Meta-Regression Analysis Fixed Effects Fixed Effects Random Effects Random Effects Variables (Weight1) (Weight2) (Weight1) (Weight2) (1) (2) (3) (4) 1.533*** 1.044** 1.093*** 1.246*** SE (4.682) (2.161) (4.046) (3.738) -0.003 0.005 -0.008* -0.007 ADL (-0.924) (0.666) (-1.848) (-0.819) -0.010*** -0.009*** -0.020*** -0.025*** Disease (-4.312) (-3.756) (-5.064) (-4.461) -0.021*** Mortality ------(-3.728) -0.006 -0.014** Obesity ------(-1.267) (-2.033) -0.012 Depression ------(-1.421) 0.012*** 0.019*** 0.012*** SelfReported ---- (4.484) (3.307) (3.318) 1.365 0.010 Secondary ------(1.362) (1.280) 0.419* Tertiary ------(1.869) 0.007** 0.018*** 0.023*** 0.012** Education attainment (2.011) (2.698) (2.839) (2.155) -0.002*** -0.002* -0.002 NumberofEducvariables ---- (-2.952) (-1.892) (-0.686) -0.006* -0.015** 0.009 EastAsiaPacific ---- (-1.727) (-2.604) (-1.656) 0.005 -0.008 Europe ------(1.338) (-1.439) 0.012** 0.018*** 0.018 Other country ---- (2.588) (3.679) (1.318) -0.008 0.014 -0.012 -0.012 Individual data (-1.130) (1.445) (-1.074) (-0.998) -0.007* -0.005 Female sample ------(-1.914) (-1.167) -0.010*** -0.014*** -0.010** Male sample ---- (-3.568) (-3.968) (-2.299) -0.024*** -0.034*** -0.021*** Other sample ---- (-5.714) (-7.756) (-3.220) -0.007** -0.008** -0.004 Age ---- (-2.527) (-2.046) (-0.990)

0.012*** 0.005 gender ------(2.918) (1.365) -0.003 -0.011* Race ------(-0.645) (-1.736) -0.003 -0.004 Marital Status ------(-0.999) (-0.943) 0.010* -0.011*** -0.013** Occupation ---- (1.807) (-2.821) (-2.035) -0.026*** -0.037*** -0.028*** -0.032*** Endogeneity (-7.421) (-2.888) (-6.267) (-6.827) -0.020*** -0.044*** -0.001** tCalculatedbyCI ---- (-4.023) (-4.241) (-2.280) 0.259** 0.183 SENonspherical ------(2.137) (1.189) -0.014*** -0.036*** -0.000 Economics Journal ---- (-3.189) (-3.883) (-1.124) -0.017*** -0.033*** -0.001*** Sociology Journal ---- (-2.729) (-3.771) (-2.829) Development & Population -0.018 0.003** 0.002 ---- Journal (-1.511) (2.078) (0.996) -0.028*** -0.004* Science Journal ------(-3.672) (-1.807) 0.005*** 0.007*** 0.002 0.003** Journal rank (5.161) (6.727) (1.486) (2.491) -0.001** -0.002** -0.001*** -0.001* Pub Year (-2.551) (-2.120) (-2.650) (-1.802) 1.615** 3.123** 2.393*** 1.622* Constant (2.590) (2.136) (2.679) (1.859) Observations 4,671 4,671 4,671 4,671 R-squared 0.430 0.555 0.459 0.506

NOTE: The results are estimated by the Weighted Least Squares (WLS) with cluster robust standard errors on the basis of equation (7). The coefficient is reported in the top and the associated t-statistics is reported in parentheses below. Two variables (SE, endogeneity) were locked into each regression specification. Backwards stepwise procedure was employed to choose the best set of control variables that minimize the Bayes Information Criterion. * p<0.1, ** p<0.05, *** p<0.01.

TABLE 10 Robustness Checks (Z-transformed PCCs and Standard Errors) Fixed Effects Fixed Effects Random Effects Random Effects Statistics (Weight1) (Weight2) (Weight1) (Weight2) FAT-PET 0.008* 0.011 0.012*** 0.015*** PET (𝛼 ) (1.81) (1.6) (4.0) (3.99) 1.61*** 1.953** 1.136*** 1.418*** FAT (𝛼 ) (2.76) (2.32) (4.27) (4.46) Observations 4,671 4,671 4,671 4,671 PEESE 0.011*** 0.013** 0.021*** 0.024*** Precision (𝛼 ) (3.16) (2.25) (7.98) (8.69) 48.48*** 53.51*** 19.69*** 26.48*** Bias (𝛼 ) (3.51) (3.44) (2.77) (3.76) Observations 4,671 4,671 4,671 4,671 NOTE: All the results are estimated by Weighted Least Squares (WLS) with cluster robust standard errors as described in equation (6) and equation (7). T-statistics is reported in parenthesis. * p<0.1, ** p<0.05, *** p<0.01.

TABLE 11 Robustness Checks (Full Sample) Fixed Effects Fixed Effects Random Effects Random Effects Statistics (Weight1) (Weight2) (Weight1) (Weight2) FAT-PET 0.008* 0.01 0.013*** 0.005 PET (𝛼 ) (1.76) (1.6) (3.18) (0.69) 1.857*** 2.362*** 1.326*** 2.645*** FAT (𝛼 ) (3.13) (2.82) (4.02) (3.61) Observations 4,718 4,718 4,718 4,718 PEESE 0.011*** 0.013** 0.024*** 0.026*** Precision (𝛼 ) (3.23) (2.27) (7.62) (8.54) 52.14*** 61.46*** 18.2*** 38.8*** Bias (𝛼 ) (3.81) (4.21) (2.53) (4.04) Observations 4,718 4,718 4,718 4,718 NOTE: All the results are estimated by Weighted Least Squares (WLS) with cluster robust standard errors as described in equation (6) and equation (7). T-statistics is reported in parenthesis. * p<0.1, ** p<0.05, *** p<0.01.

FIGURE 1 Histogram Distribution of t-Statistics

Distribution of t-statistics Obs. Percent t < -2.00 169 3.62 -2.00 ≤ t ≤ 2.00 2,319 49.65 t > 2.00 2,183 46.74 Total 4,671 100

FIGURE 2 Distribution of Partial Correlation Coefficient (PCC)

FIGURE 3 Distribution of PCC Values by Health Measures

A. Physical Health

B. Mental Health

C. General Health

FIGURE 4

Funnel Plots

A. Individual Estimates

B. Mean Study Estimates

APPENDIX A

PRISMA Flow Diagram for Inclusion in the Meta-Analysis

Records identified through database search (n=474)

Identification Identification Records after duplicates removed Studies excluded on title/abstract(n=161) (n=363) ·106 studies on health education ·4 studies on the effect of health on education ·51 studies on the intergeneration effect of education on health Records screened(n=202) Screeing Full articles excluded(n=25) ·9 Insufficient information to extract an effect ·7 path analysis ·9 Interaction terms or quadratic terms in Studies assessed for eligibility education (n =177)

Eligibility

Studies included pure theory or qualitative synthesis (n=70)

Studies included in meta-analysis (n=107) Included

APPENDIX B

List of Studies

ID Authors PubYear Journal Name Estimates American Economic Journal: Applied 1 Meghir et al. 2018 Economics 10 2 Clark & Royer 2013 American Economic Review 18 3 Oreopoulos 2006 American Economic Review 8 4 Madsen et al. 2010 American Journal of Epidemiology 93 5 Ross & Wu 1995 American Sociological Review 13 6 Becchetti et al. 2017 Applied Economics 72 7 Zhong 2015 Applied Economics 12 8 Cook 2018 Applied economics letters 4 9 Zhong 2016 China Economic Review 50 10 Xie & Mo 2014 China Economic Review 79 11 Castro 2012 Demographic Research 42 12 Ross & Mirowsky 1999 Demography 24 13 Lundborg et al. 2016 Demography 50 14 Atella & Kopinska 2014 Demography 60 15 Behrman et al. 2011 Demography 20 16 Grossman 2008 Eastern Economic Journal 2 17 Mazumder 2008 Economics Perspectives 116 18 Ma et al. 2018 Economics and Human Biology 18 19 Amin et al. 2018 Economics and Human Biology 12 20 Dursun et al. 2018 Economics and Human Biology 238 21 Kim 2016 Economics and Human Biology 41 22 Arendt 2005 Economics of Education Review 14 23 James 2015 Economics of Education Review 39 24 Groot & Brink 2007 Economics of Education Review 74 25 Jamison et al. 2007 Economics of Education Review 6 26 Silles 2009 Economics of Education Review 32 27 Amin 2013 Economics of Education Review 57 28 Adams 2002 Education Economics 25 29 Edwards 2015 Education Economics 23 30 Hirokawa et al. 2006 European journal of epidemiology 60 31 Brunello et al 2015 Health Economics 16 32 Dee & Sievertsen 2018 Health Economics 10 Kiuila & 33 Mieszkowski 2007 Health Economics 60

34 Seo & Senauer 2011 Health Economics 10 35 Van der pol 2011 Health Economics 53 36 Auld & Sidhu 2005 Health Economics 12 Fujiwara & 37 Kawachi 2009 International Journal of Epidemiology 53 38 Bann et al. 2016 International Journal of Epidemiology 30 International Journal of Health 39 Ayyagari et al. 2011 Economics and Management 16 40 Bai & Li 2018 Journal of Economic analysis & policy 20 41 Braakmann 2011 Journal of Health Economics 45 42 Albouy 2009 Journal of Health Economics 11 43 Kemptner et al. 2011 Journal of Health Economics 80 44 Kemna 1987 Journal of Health Economics 24 45 Cutler et al. 2011 Journal of Health Economics 20 46 Grimard et al. 2010 Journal of Health Economics 24 47 Buckles et al 2016 Journal of Health Economics 25 48 Black et al. 2015 Journal of Health Economics 45 49 Palme & Simeonova 2015 Journal of Health Economics 16 50 Webbink et al. 2010 Journal of Health Economics 82 51 Powdthavee 2010 Journal of Human Capital 22 52 Hardarson et al. 2001 Journal of Internal Medicine 24 53 Brunello et al. 2013 Journal of Labor Economics 14 54 Oreopoulos 2007 Journal of Public Economics 12 55 Loundborg 2013 Journal of population economics 63 56 Jurges et al. 2013 Journal of population economics 40 57 Leuven et al. 2016 Labour Economics 20 58 Devaux et al. 2011 OECD Journal:Economic Studies 170 59 Lager &Torssander 2012 PNAS 59 60 Lutz & Kebede 2018 Population and Development Review 1 61 Böckerman et al. 2017 Preventive Medicine 6 62 Lleras-muney 2005 Review of Economics Studies 26 63 Gerdtham et al. 2016 Scand. J. of Economics 35 64 Lynch & Hippel 2016 Social Science & Medicine 12 65 Bijwaard et al. 2017 Social Science & Medicine 88 Bockerman & 66 Maczulskij 2016 Social Science & Medicine 48 67 Benson et al. 2018 Social Science & Medicine 40 68 Zhang 2010 Social Science & Medicine 30 69 Zajzcova et al. 2012 Social Science & Medicine 88 70 Huang & Zhou 2013 Social Science & Medicine 23 Knesebeck & 71 Dragano 2006 Social Science & Medicine 140 72 Luo et al. 2015 Social Science & Medicine 84 73 Montez et al. 2018 Social Science & Medicine 20

74 Fletcher 2015 Social Science & Medicine 44 75 Mazzonna 2014 Social Science & Medicine 18 76 Denney et al. 2010 Social Science Research 96 77 Rogers et al. 2013 Social Science Research 168 78 Parinduri et al. 2016 The Journal of Development studies 63 79 Berger & Leigh 1989 The Journal of Human Resources 4 80 Asghar et al. 2009 The Pakistan Development Review 20 81 Arkes 2003 RAND Working Paper 2 82 Barcellos et al. 2018 CESR-SCHAEFFER working paper 8 83 Bijwaard et al. 2016 IZA working paper 36 84 Braakmann 2010 University of Lüneburg working paper 6 85 Braakmann 2010 University of Lüneburg working paper 132 86 Brunello et al 2009 working paper 37 87 Buckles et al 2012 IZA working paper 32 88 Cawley & Choi 2015 NBER working paper 54 89 Cesur et al 2014 NBER working paper 89 90 Chevalier&Feinstein 2006 IZA working paper 200 91 Cipollone & Rosolia 2011 Temi di discussione working paper 32 92 Clark & Royer 2010 NBER working paper 28 Cutler & Lleras- 93 Muney 2006 NBER working paper 51 94 Fonseca et al. 2018 working paper 24 95 Fonseca et al. 2011 RAND working paper 24 96 Grabner 2009 working paper 176 97 Huang 2015 IZA working paper 24 98 Janke et al. 2018 IZA working paper 36 99 Lacroix 2018 working paper 19 100 Lillard & Molloy 2007 working paper 5 101 Lundborg 2008 Tinbergen Institute Discussion Paper 30 102 Meghir et al. 2012 IZA working paper 14 103 Albarran et al. 2017 working paper 15 104 Quis & Reif 2017 conference paper 72 Schneider & 105 Schneider 2006 working paper 3 106 Spasojevic 2010 books 15 107 Deaton & Paxson 1999 NBER working paper 78